The global memory shortage is not just about supply and demand. Geopolitics is actively redrawing the map of who gets to make what, and where. U.S. export controls are blocking China from accessing advanced memory technology. Proposed tariffs on Korean and Taiwanese chips threaten to jack up costs for everyone else. And China’s biggest DRAM maker, ChangXin Memory Technologies (CXMT), is ramping up conventional DRAM production but remains years away from manufacturing the high-bandwidth memory (HBM) that AI actually needs.
Here is the thing you need to understand: coverage of Chinese memory production routinely conflates conventional DRAM with HBM. They are not the same thing. This article separates what CXMT can realistically deliver from what it cannot, and gives you the geopolitical context for the global AI memory shortage without requiring a trade policy degree.
What Is China’s Actual Share of Global DRAM Production?
CXMT is based in Hefei, produces DDR5 and LPDDR5 at scale, and is now the world’s fourth-largest DRAM manufacturer — holding roughly 15% of global output.
But here is what matters: all of that is conventional DRAM. CXMT does not produce HBM, which is the memory type in shortest supply for AI workloads. Technologically, it sits about two process generations behind Samsung and SK Hynix.
Then there is Yangtze Memory Technologies (YMTC), China’s NAND flash leader, which is constructing a third fab in Wuhan with around half its planned output earmarked for DRAM. But YMTC’s DRAM ambitions are even further from commercial scale than CXMT’s HBM plans.
HP, Dell, Acer, and Asus are actively qualifying CXMT DRAM for non-U.S. devices. That is a legitimisation signal, but HP will reportedly only put CXMT chips in devices sold outside the United States — a deliberate compliance firewall. So CXMT is a real player, but its relevance to the Samsung/SK Hynix/Micron competitive position is confined to conventional DRAM in specific markets.
Do U.S. Export Controls Make the Memory Shortage Better or Worse?
Short answer: worse.
U.S. Bureau of Industry and Security (BIS) export controls restrict the transfer of advanced semiconductor technology to China — including HBM stacking processes and packaging equipment from ASML, KLA, and Applied Materials. The rules limit tool sales for sub-18nm DRAM processes and target advanced packaging tech relevant to HBM.
The practical effect is a technology ceiling. CXMT simply cannot get the equipment it needs to manufacture HBM at competitive yield. And because Chinese AI companies like Huawei cannot access Korean or U.S. HBM either, domestic demand concentrates on whatever CXMT can produce. CXMT’s growth stays inside China. It does not relieve global HBM shortages one bit.
It is worth understanding the distinction between export controls and tariffs. Controls restrict China’s ability to produce competitive alternatives. Tariffs increase the cost of allied supply. Both make things worse for enterprise buyers, but the question of whether Chinese growth changes the shortage resolution timeline depends entirely on which lever you are looking at.
CXMT is also on the U.S. Department of Defense “Chinese military company” list. It is not banned outright, but the designation creates compliance review requirements for any Western company even considering CXMT as a supplier.
One more thing worth knowing. Samsung and SK Hynix both operate large fabs inside China that are now frozen in capability — running, but unable to upgrade. As Chinese domestic production expands, both companies may eventually exit the Chinese market entirely. That is an underappreciated supply variable that most analyses skip over.
Can CXMT and YMTC Provide Meaningful Supply Relief?
In the near term (2026–2027), CXMT provides some conventional DRAM relief for consumer devices and domestic Chinese buyers. It is targeting HBM3 by end of 2026 and has delivered HBM3 samples to Huawei. But samples and volume production are very different things.
The challenge is mastering through-silicon via (TSV) stacking — bonding multiple DRAM dies into a single HBM package, where a single misstep ruins the whole stack. Chinese equipment makers like Naura Technology are developing domestic tooling, but none of it is confirmed ready for mass production.
So do not plan your procurement around Chinese HBM. It is not going to deliver at volume in the 2026–2027 timeframe. Looking further out to 2028 and beyond, Chinese HBM at scale is plausible but not guaranteed. Counterpoint Research expects Korean firms to maintain dominance.
What Does Supply Chain Diversification Actually Mean for the Shortage Timeline?
New fab investments are underway. Micron is building in Taiwan and New York. SK Hynix is investing in Indiana. Expansions are happening in Japan and Singapore. These are partly enabled by the U.S. CHIPS Act, and Taiwan has committed approximately $318 million in subsidies to Micron for HBM R&D. For full context on how the memory shortage is reshaping the entire tech sector, see the global AI memory shortage overview.
But new fabs come online in the 2027–2030 window. For the current shortage, diversification is a strategic hedge, not a supply solution.
Then there is the tariff threat. U.S. Commerce Secretary Howard Lutnick has warned Samsung and SK Hynix they could face 100% tariffs. His exact words at Micron’s New York fab groundbreaking: “Everyone who wants to build memory has two choices: They can pay a 100% tariff, or they can build in America.”
And here is the part that gets missed in most coverage: long-term agreements (LTAs) between hyperscalers and memory makers lock up existing capacity for AI workloads. New fab output may already be committed before it reaches the open market. Diversification does not put more memory on the spot market for enterprise buyers anytime soon.
Is There a Scenario Where Memory Demand Collapses Before New Supply Arrives?
This is the question that matters most for capital planning.
Deloitte’s 2026 semiconductor outlook flags that the industry may need to prepare for a demand correction. If enterprises cannot demonstrate returns on AI investment, the hyperscaler capex driving HBM demand could slow down. The DeepSeek efficiency narrative fed into this — if models get more efficient, maybe memory demand falls.
But committed spending tells a different story. Microsoft, Google, Amazon, and Meta have committed record AI infrastructure spending. Stargate — the $500B OpenAI/SoftBank/Oracle joint venture — represents demand spanning years. Hyperscalers are placing open-ended memory orders, accepting any HBM volume at any price. That does not look like demand about to fall off a cliff.
The reconciliation is inference scaling. More efficient models get deployed more broadly. AI workloads have expanded from training into inference and decision-making — and each stage adds memory demand at the system level. Efficiency per model can actually increase total consumption. It is a risk scenario, not a forecast — but it is worth stress-testing your procurement assumptions against.
What Does All of This Mean for Enterprise Memory Sourcing?
Chinese DRAM provides minimal near-term relief for Western enterprise buyers. CXMT does not serve the global spot market at scale, and the DoD blacklist creates compliance requirements most procurement teams are not set up for. Qualifying CXMT’s chips does not mean OEMs will automatically order from them. It is a contingency measure for non-U.S. markets. Enterprise-grade supply for Western buyers remains controlled by Samsung, SK Hynix, and Micron — and understanding how those three makers locked up supply with hyperscalers is essential context for any sourcing negotiation.
So what should you actually do? Check whether your supply chain has indirect CXMT exposure through contract manufacturers. Build sourcing strategies that account for export control tightening and tariff escalation as separate scenarios. And secure supply commitments with established suppliers before the tariff situation resolves — waiting carries price spike risk regardless of which geopolitical path unfolds.
The geopolitical layer of the broader memory supply crisis does not change your near-term options much. But it changes the medium-term calculus considerably, and it is worth understanding before you lock in multi-year sourcing strategies.
FAQ
Does CXMT sell memory directly to enterprise buyers outside China?
Not at meaningful scale. CXMT primarily serves the domestic Chinese market. Enterprise-grade memory for Western markets remains dominated by Samsung, SK Hynix, and Micron.
Is it legal to buy CXMT memory for products sold in the United States?
CXMT is on the U.S. DoD “Chinese military company” list, which restricts certain transactions. It is not an outright ban, but you should conduct a formal export control compliance review before incorporating CXMT components into any U.S.-destined product.
When will China be able to manufacture its own HBM?
CXMT is targeting HBM3 production by end of 2026, but volume production at competitive yield is more realistically a 2028+ prospect. The technology gap in TSV stacking and domestic equipment maturity are the primary constraints.
What is the difference between U.S. export controls and tariffs on memory chips?
Export controls block China’s access to manufacturing equipment. Tariffs increase the cost of memory from allied nations like South Korea and Taiwan. They are distinct policy levers with different supply-side effects, but both make the shortage worse.
Will the proposed 100% tariff on Korean and Taiwanese memory actually happen?
As of early 2026, it has not been enacted. But the threat alone is already influencing procurement decisions and pricing expectations. Plan for multiple tariff scenarios.
How does the DeepSeek efficiency narrative affect memory demand?
More efficient models reduce memory demand per model, but they get deployed more broadly. Inference scaling tends to sustain or increase total memory consumption at the aggregate level.
What happens to Samsung and SK Hynix fabs inside China?
They continue operating at their current technology level but cannot upgrade due to export control rules. The long-term future of these assets is unclear and represents an underappreciated variable in supply projections.
Can supply chain diversification solve the memory shortage?
Not within the 2026–2027 window. New fabs operate on 3–5 year construction and ramp timelines. Diversification reduces geographic concentration risk long term but does not provide near-term relief.
What is the CHIPS Act’s role in memory supply?
The U.S. CHIPS Act subsidises domestic semiconductor manufacturing, supporting fab investments by Micron and SK Hynix on U.S. soil. It accelerates geographic diversification but does not change the timeline for new capacity coming online.
Should enterprise procurement teams worry about indirect CXMT exposure?
Yes. As OEMs qualify CXMT memory, components sourced through contract manufacturers may include CXMT DRAM. Audit your supply chain for indirect exposure, particularly for products with U.S. market distribution.
